Arista Etherlink 800G Multi-Mode AOC

The Arista Etherlink 800G Multi-Mode AOC is a high-density active optical cable purpose-built for the bandwidth and latency demands of modern AI and machine learning infrastructure. Leveraging an 8x100G PAM4 electrical interface on each end, the cable integrates VCSEL-based optical transceivers with onboard DSP-assisted signal conditioning to deliver error-free transmission across multi-mode fiber spans up to 100 meters. The active optics architecture eliminates the passive copper limitations that constrain bandwidth density in large-scale GPU-to-GPU and GPU-to-switch topologies, making it an ideal interconnect fabric for NVIDIA DGX, AMD Instinct, and custom AI accelerator clusters operating over Arista 7800R3 and 7060X series spine-leaf networks.

Designed in alignment with IEEE 802.3df 800GbE and OSFP-compliant mechanical specifications, each end of the AOC terminates in an OSFP connector housing that is hot-pluggable and backward-compatible with 400G port modes via gearbox negotiation. The onboard retimer circuitry compensates for signal integrity degradation across the fiber span, maintaining BER below 10⁻¹⁵ post-FEC without requiring host-side equalization overhead. This design philosophy directly reduces CPU and NIC offload burden on connected accelerator nodes, freeing compute resources for inference and training workloads rather than link management.

Launched in Q4 2025, the Arista Etherlink 800G AOC targets hyperscale AI factories, high-performance computing facilities, and colocation operators deploying next-generation GPU clusters. Its power envelope of under 5W per end is among the lowest in the 800G AOC segment, translating to meaningful PUE improvements at rack scale when hundreds of interconnects are deployed simultaneously. The cable ships in standardized lengths from 1 meter to 30 meters, with OM4 multi-mode fiber ensuring compatibility with existing data center fiber plant investments.

Ideal for

• GPU-to-GPU direct interconnect within large-scale AI training clusters using 8-way or 16-way NVLink-alternative topologies over structured OM4 fiber runs
• Spine-leaf fabric cabling between Arista 7800R3 series core switches and top-of-rack leaf switches in hyperscale AI data centers requiring sub-100ns hop latency
• High-density compute-to-storage interconnects linking GPU servers to all-flash NVMe-oF storage arrays where InfiniBand or Ethernet at 800G is required for checkpoint and dataset streaming
• Colocation cross-connect within high-density AI pods where copper DAC cables are impractical beyond 3 meters and coherent optics are cost-prohibitive for intra-rack links
• Disaggregated AI inference platforms connecting pools of accelerator trays to centralized load-balancing switches in carrier-grade AI-as-a-Service deployments
• HPC simulation environments requiring deterministic low-latency MPI traffic across fat-tree topologies where consistent 800G per-link throughput is critical to parallel workload completion times

Technical specifications

Available from Omnixon Global. Submit an RFQ and our team will confirm configuration and availability for your order.